Nothing is perfect in the early stages - but you seed the "nude"...

Nothing is perfect in the early stages - but you seed the "nude" mice (significantly compromised immune system) with a human cancer, and then the only agent that should be statistically relevant is the agent itself that is used to "treat" the mice.
So you can see in a living organism that is very similar to a human in many ways, and what happens to the cancer with the different treatments is comparable.

Always a risk that in humans it will work out differently. But these tests do have a good track record for predicting success in human subjects.

If I have got anything wrong here - pls let me know.